PU-MFA: Point Cloud Up-Sampling via Multi-Scale Features Attention.

Area of Science:

• Computer Vision
• 3D Data Processing
• Deep Learning

Background:

• 3D scanner technology advancements drive increased use of point clouds.
• High-quality point cloud acquisition remains costly, creating a demand for efficient generation methods.
• Deep learning offers a promising solution for point cloud up-sampling.

Purpose of the Study:

• To propose a novel deep learning method, Point cloud Up-sampling via Multi-scale Features Attention (PU-MFA), for generating high-quality point clouds.
• To address the challenge of high costs associated with obtaining high-quality point clouds.
• To leverage multi-scale features and attention mechanisms for effective point cloud refinement.

Main Methods:

• Developed PU-MFA, integrating multi-scale features and attention mechanisms within a U-Net architecture.
• Employed adaptive multi-scale feature utilization to refine global features.
• Conducted experiments on both synthetic (PU-GAN) and real-scanned (KITTI) datasets.

Main Results:

• PU-MFA demonstrated superior performance in generating high-quality dense point sets compared to state-of-the-art methods.
• Quantitative and qualitative evaluations confirmed the effectiveness of the proposed method.
• Visualizations of attention maps illustrated the impact of multi-scale features.

Conclusions:

• The proposed PU-MFA method effectively addresses the need for high-quality point cloud generation.
• The integration of multi-scale features and attention mechanisms proves beneficial for point cloud up-sampling.
• PU-MFA represents a significant advancement in efficient and cost-effective point cloud enhancement.